Verbosity Tradeoffs and the Impact of Scale on the Faithfulness of LLM Self-Explanations

Larger AI models provide more faithful explanations: The research found a consistent scaling trend showing that bigger and more capable language models produce explanations that more accurately reflect their actual decision-making processes, suggesting that scale itself may improve trustworthiness.

There's a tradeoff between detailed and concise explanations: The study identified tensions between verbosity (longer explanations) and comprehensiveness (complete explanations), meaning models often face a choice between giving brief answers or thorough ones—with implications for how much users can actually trust what they're told.

Plausible-sounding explanations can be misleading: Even when LLM explanations sound convincing to humans, they may not faithfully represent what actually drove the model's decision, highlighting a key challenge in responsible AI deployment where users might be misled by seemingly reasonable explanations.

Current evaluation metrics have limitations: The paper demonstrates that existing metrics for assessing explanation quality can be "gamed" or manipulated, and proposes new metrics (phi-CCT and F-AUROC) to better measure whether explanations are genuinely faithful rather than just appearing credible.

Testing faithfulness requires rigorous methods: The research emphasizes that determining whether AI explanations are trustworthy requires systematic counterfactual testing across diverse models, not just human judgment, underscoring the need for technical rigor in responsible AI evaluation.